D-Link switch and ecc-memory.
tegner at nada.kth.se
Tue Jan 16 15:38:35 EST 2001
>> we have a small system, eight nodes with D-Link DFE-530 TX nics and a
>> D-Link DES-3225G switch.
>> After an update of PROM code and Runtime Code on the switch the
>> situation was improved slightly (i.e, we got acceptable performance
>> using three nodes even if the switch was not restarted).
>Are you implying that more than three nodes causes additional problem
Yes, initially, before the upgrade, communications between any two nodes
over the switch would typically "die" when the package size was larger
than around 5800 bits. After the upgrade of the switch this "Mbits
packet size curve" (obtained by netpipe) looked OK. So I figured problem
solved, and started testing the speedup of the cluster, and this also
reasonably up to three nodes, but after that it was really lousy. If I
restarted the switch again that problem went away.
>> It doesn't seem to matter if we specify full-duplex for the nics in
>> conf.modules, and I wonder if anyone else have experienced similar
>Use 'cat /proc/net/dev' to check for errors. You should have very few
>errors in a properly working cluster network. Specifically, there
>should never be "frame" or "Tx carrier" errors.
I checked, no errors, even without restarting the switch.
>The 'mii-diag' program is useful for verifying that the network
>configuration is correct. All MII transcievers provide basic
>information such as speed, duplex and negotiated flow control. Some
>transceiver provide additional information such as a symbol error count
>and noise margins.
Will try this next, and many thanks for all hints!
Beowulf mailing list
Beowulf at beowulf.org
More information about the Beowulf